Not sure why you had to add the (attempted) qualifier. He started a company and is selling a box. That makes him a merchant. How successful that venture is, is a different question, but he absolutely is a merchant in this arena.
A slight defense of any qualifiers, mine or others. I am a sole developer, who formed a "box" as you put it, to house my assets. I'm the sole director of it. That doesn't mean my box is at war with users. To be clear, a merchant is a middleman and some merchants have taken the goods produced by the builder and built a middleman moat. Cloudflare, Anthropic, and the list goes on. Geohot doesn't do this. He builds and sells what he builds. That makes him more a craftsman, not a merchant. Merchants market other's work and profit. Craftsman, or builders, build and grow their customer base organically.
I feel like you're splitting hairs. Geohot is/was selling shovels during a gold rush, full stop.
Maybe he had a recent change of heart, but his public actions (including Elon stuff) clearly put him in the "monetize hype" camp rather than the "quietly build" camp.
>One, this constant bullshit about some window closing, or the perpetual underclass, or falling hopelessly behind. This is negative valence hype, not only is it not true, it’s mostly designed to make you feel bad about yourself and move to shitty San Francisco where everything really does suck like how these people claim.
It's possible to use LLMs without logging onto twitter to be exposed to the people spouting off about a "perpetual underclass." I love the internet, but it really feels like (now more than ever) you have to be intentional about what sites you visit.
"Permanent underclass" is the notion that people who get involved at the ground floor will essentially get infinite wealth relative to the ones who don't. It's a little goofy, but more of the capitalism you'd expect from today's X than the communism you're imagining in yesterday's Twitter.
Agreed. There's sort of this spiteful anti-hype here that I find very offputting, and ultimately I think it's because a lot of folks are going out and encountering opinions I never see. I hear wild conspiracy theories about data centers and the financials of involved companies that make their way to me from bluesky or instagram, often through here, but never the unstoppable tide of hype that people are allegedly[1] railing against. I do read Scott Alexander, but he's a lot more reserved than people make him out to be on this.
[1] Allegedly because I have no firsthand experience, not to imply doubt.
Agreed, but I do think this is a wholly different kind of hype. With crypto currencies it was the promise of modernizing value exchange, with some zealots promising the end of traditional currency.
With this, I’m hearing (from supposedly reputable publications, in addition to random people) that this is going to end knowledge work in general and take out a large percentage of the world’s labor force. I’m being told to pick up a trade, and that the career I have and the knowledge I’ve gained is now worthless.
The worst part seems to be that it’s pretty much impossible to quantify any kind of impact these tools will have until after the impact is actually felt. We’ve been in limbo while the tech sector is just rotting.
It is great for video games because it gives you fresh noobs to pwn instead of having to face the sweats. I can’t really play the old fps games I like anymore not because they are dead, they are very much alive, but everyone still playing has been playing nonstop for years now and are way above me in skill.
There are many things to be critical about but shoehorning an entire metro into the echo-chamber you're supposedly beyond yet can't help but orient your entire world view as the anti-SF-tech-bro all while running a startup and discussing AI on HN.
TLDR: SF is more than Paul Graham worship parties.
the vast majority of the target audience of this blog post would only consider moving to SF because of the tech scene. This isn't a mountain biking or asian food blog.
> What I don’t like is two things. One, this constant bullshit about some window closing, or the perpetual underclass, or falling hopelessly behind.
> And two, this strawman jump from, oh hey, it’s a fancy autocomplete, smart compiler, better search engine, to it’s gonna like own the whole light cone bro like if you aren’t in SF and at the right parties there’s gonna be like a flash of light in the sky one day and you’re not even gonna know what happened but everything just Changed.
Haha, OP has a way with words.
In a way, both these emotional extremes (FOMO & the singularity) are just tools being used to continue driving the massive CapEx behind LLM improvement. Hate to love it? Love to hate it?
How to you love this stuff so hard? I could newer love any ai generated music, book or artwork. Anything ai gemerated i have ever seem or heard was either disgustingly slop or indistinguishable from something else which was real. It‘s a like finding a cool track only to discover it‘s a lazy bootleg.
Yeah but it was only like 2 years ago that artists were arguing this on the basis that AI-gen images would consistently mangle hands
Now we're at a point where that never happens, and where lipsync is almost a completely solved problem
If the issue here is simply that the quality is bad, one has to contend with the fact that it is undoubtedly exponentially improving and there's no reason we should expect that improvement to stop
I also don't have any interest in consuming AI generated art, but the same criticisms were levied at computer graphics and if we're comparing to CGI we'd be at the late 1970s in terms of nascency
An LLM cannot make art because it isn't human. It can make "art like artifacts". Art involves one human communicating some emotional experience to another human, LLMs cannot experience human emotion, so they cannot make art.
The process of making art is not a subset of hill climbing optimisation algorithms.
By this definition most of our culture isn't art because it's made not to "communicate some emotional experience" but to induce one in order to make money. 90% of pop music is like that, for example. And LLMs can do that just as well - better, probably, since they know all the psychological tricks from their training data.
No, even the most cynical pop music is going to carry more emotional resonance with the average person, because art is a social contract between the Creator and audience. Even if you are keenly aware that any perceived bond with an artist is parasocial, humans get a fundamentally different experience if they think a real human was behind it.
It has nothing to do with quality. Artists that use AI are going to need to hide it because people will enjoy it less knowing it's AI. It's that simple. Maybe that will change in 15 years if a new generation is trained to believe learning a skill is stupid or embarrassing. I wouldn't rule that out, these companies are already trying to convince people not using their products is morally bad.
The natural world is full of things that area beautiful and breathtaking without having come from "human emotion".
The plumage of a peacock is beautiful and awe-inspiring beyond most human made art, and it is genuinely the result of evolutionary hill climbing on a fitness landscape
I remember watching Jurassic Park when it came out and i loved the CGI (the 5 seconds). Now with Ai we should be excited to see the new dinosaurs looking even more realistic but somehow they don’t- they look like in 1996 just not quite as good…
> where’s all this new magical software that the productivity improvements should imply?
It's running, privately, in my homelab.
I think we are entering what I call the "have it your way" era. If an open source project doesn't do exactly what you want it to do, fork it, or create a new version. It's too easy.
This makes me a bit concerned about the future of open source. Upstreaming used to be worth it, since maintaining a fork is effort too. But now the balance has shifted significantly. Especially with many projects becoming a lot stricter about contributing, and some becoming outright hostile to AI. I can't blame them. But I think the effect will be that improvements are less likely to make it back to the community as AI adoption increases.
You will likely end up in maintenance hell soon. This will likely not be much easier with AI because coding is not the hard/annoying part, it's the fact that you need to dust off every little project every time a tiny fix is needed, and that's a lot of toil in the long run.
Maybe? I ran across an old pre-LLM project of mine recently, and past me was an asshole and didn't leave a readme for future me. Meanwhile post-LLM projects at least have a readme that the LLM generated for me or my agent to read and pick up context on. Being able to ask an agent what is this repo, what's going on here? Hey just make it do this, instead of toilsomely digging in and doing it tmmyself, seems to say that might not come to pass.
There is, of course, the question of if that's making me dumber. It might be, but there are other brain training things I'm doing outside of that to force my brain to do the thing.
The fact that you're even saying this it is probably an admission that you do think it's making you dumber. Most people I know, who are honest with themselves, have admitted to me that they feel like it's making them dumber or "zombifying" them. This is also well studied already, https://arxiv.org/abs/2506.08872
LLMs are poison for the brain, I'm almost certain of it, at least when used in the way most people are using them. If you drive everywhere because you don't want to walk (but you could), you're obviously going to be physically worse off than if you walked. This is the case with llms, if you have them do all the thinking, planning and action you're going to be cognitively worse off than if you didn't use them.
Socrates was right, honestly. I would not be surprised if humans are more evolutionarily optimized for that sort of communication and thinking. Maybe socrates noticed that writers/readers vs orators/listeners were indeed generally dumber in the way we consider people riddled with short form social media induced brainrot to also be missing some mental capacity.
Widespread literacy is only a few generations old, arguably I guess. Meanwhile we’ve been speaking to eachother for longer than we’ve been humans. Oral information can be kept longer than written information too it seems. Our oldest kept information is not written down, but in folk stories such as aboriginal tales some tens of thousands of years old.
It's pretty easy to generalize this, but it doesn't match my perception. People who are using llms to do things they could have already done, but faster, probably have atrophying skill sets. People who are using these tools to accomplish significantly more difficult or complex work than they used to are absolutely finding new ways to push themselves. The problems are just much bigger.
The average Joe can easily vibe code apps that took a small startup just a few years ago. If developers are also using AI to build the same simple apps - then yeah. They're not pushing themselves hard enough, and probably not using their brains as much anymore.
This matches my experience as a statistician who used to begrudgingly write bad code when I had to. LLMs have opened up huge new possibilities for me.
No doubt there's some Gell-Mann amnesia going on, because I regularly have to correct them from doing stuff that's really dumb based on my expertise within my area of specialization. More than once I've managed to extract >3 orders of magnitude performance gains after asking them to justify why their code was so slow. Probably there's still some stupid stuff in there. But it's better than the code I would have written, and I never could have paid for a proper developer to write it.
My perception is that this guy's response to "I forgot to write a readme" was "I should limit myself to tools that do it for me" instead of instilling discipline about documentation.
"People who are using these tools to accomplish significantly more difficult or complex work than they used to are absolutely finding new ways to push themselves"
hahaha this is a miniscule amount of people.
most people do not care about their job, only to the extent it is a source of income. but they do not care about it anything more than that. and they shouldnt either!
I don't see average Joes vibe coding apps that before took a small startup, I see a lot of cheap talk around that idea, but no receipts ... I see some incredibly simple chrome extensions and the such out there, with their api keys hard coded in their client side code lol
The worrying thing is these are people who nominally seem smart - what we are seeing is 'smart' is not what we think/was. Smart is the ability to identify an object which is harmful and a) be disciplined about its use b) become a better person who doesnt need to expose themselves to said object.
I actually I had one fella who was very distressed - pouring out all his stresses to a bot. Then I reminded him the bot is literally designed to tend to his nees - not to his benefit - but to the provider of the good whom will manipulate said user later on. He then immediately deleted his accounts.
Many are not aware of whats going on around them - this is very concerning.
Sometime a wrong and misleading README done more harm than good.
It is not that rare to see LLM waste hours on a wrong path because a misleading line in README. Even worse, they can't learn. Spawn a subagent and it repeat the same error again
Someone on my team started using an LLM to write all his readme files (he used to not write them at all).
So far, 100% of them have been wrong. I read them, my spidey senses think that what it says doesn’t match his style. I look at the code to find the variables the readme mentions doesn’t exist anywhere in the codebase. I then reach out to him about it, where it says it was written by AI and he will go back and write it for real.
He says it’s better than nothing, but agree with you that it does more harm than good. I wasted my time reading slop. I wasted more time validating the slop. I wasted even more time with a conversation about it. Now he’s spending time re-writing something that he could have written faster and better when he was actually writing the code and it was fresh in his mind. Meanwhile, I’m either blocked waiting for him, or I need to spend my time trying to understand the minutiae of his code so I can integrate it into mine.
alternatively, you might end up in 'good enough heaven' and not have to touch it for a decade because, you know, it does exactly as you need and you're not google, microsoft, openAI or antrhopic.
I'd bet there's far more 'good enoughs' than anything else out there. One of the reasons microsoft office is constantly churning subscription, etc is because they solved good enough decades ago and need to justify valuations that just don't matter for most of their user's use cases.
Not everyone is a software developer having to churn out the 101th SaaS that's just because some MBA refuses to hire a dev.
Seems to me this would get easier or harder depending on how you write the code. Like if you write the code in something standard and unchanging like POSIX shell scripts or C99 or ES5 javascript, at least the ecosystem won't change out from under you. If you use rust or python or a bunch of node.js dependencies then you might have to edit the project just to keep up with ecosystem changes.
yeah I had this happen to me. Except when I go to maintain it, now cursor/claude are good enough to essentially handle it on their own, so it turns out to be very low effort to maintain.
"This new tool allows for writing all this code ..... but every person and company, in unison, in a grand conspiracy, all decided to only write private software with it that they aren't releasing to the public in any way"
Doesn't have to be "every person and company, in unison, in a grand conspiracy" and other such strawmen.
We could try steelmaning this argument instead: it's enough that most big companies who would otherwise have incentives to contribute.
Before FOSS got in fashion, around the early 2000s, most commercial companies wouldn't touch it as contributors and were openly avert to it, and to open sourcing their stuff. This can be the case again.
Creating a fork of an active project only makes sense if you are its sole user (of the fork) and you really need exactly the modification you've been dreaming of.
I have seen so many unnecessary forks of popular projects that I think it's better to stick with the original, even if that means it won't be perfect.
In the old world, this was because keeping in sync with upstream was hard. In the new world, it takes an hour. And because you're the only user, you can test in prod. Makes the whole thing faster. I have lots of forked and family-only software. Some are abandoned upstream etc.
As cost to software goes to zero, these things become easily possible. In the past, I'd only fork top-quality software (things like `xsv` etc. which is easy to edit. These days even complex PHP software I fork with little trouble.
With lots of software, the value is in the data model and algorithm choices. Sometimes I even just point Claude Code / Codex at an open-source thing I want to vendor some functionality into my personal setup with and it gives me what I want. The hard part for me is modeling the data well. That takes experience with encountering things and it's hard to replicate the edges. LLMs often don't get the rough bits right. But someone else's hard work usually has accounted for this.
The law of conservation of energy also applies to software. If the price of software approaches zero, it is offset by the time and tokens required to modify and maintain it. Price, time, tokens are simply different expressions of energy.
That doesn’t seem right. The time required to maintain software is less than before. Don’t know about tokens. I suppose they were zero previously and are non-zero now but this does not seem to be a conservation law because as models get better the number of tokens required has become smaller too.
As soon as we started unironically calling LLMs "AI" we went down the hype path. That has plenty of downsides, like stressing out the entire world and attracting cryptocurrency bros, but also the major upside massive of funding/acceleration.
So far, all we have is more software running on computers. It's powerful, and it's amazing, but it's not magic.
Calling it "AI" was possibly a net-negative but we don't know yet.
"It's powerful, and it's amazing, but it's not magic"
But since its creators and as of my knowledge everyone else totally did not see it coming, that you can now give a vague prompt full of spelling errors - and get returned a working program - I would say it is pretty close to magic (as in we don't really understand why it works so good).
I also don't see how you cannot call it AI. Especially since simple chess engines and alike were called AI long ago. So it is not general strong AI and has no consciousness and no mind and is pretty dumb too often - but the general concept - getting from a some vague text to a working program has some connection to intelligence to me.
One of the lesser, but still underdiscussed ramifications is that I think it has limited the public's ability to comprehend the Yann LeCunn argument, that genuine AI is likely possible but that LLMs and transformers are a dead end and we need to explore different modalities
At least for me, the jump in productivity has resulted in building stripped down one-off software for my highly specific use-cases.
You can use an LLM to create anything but you still need to know what it is that you're building, and you need to think through how everything should work or the LLM will just fill it with sausage. You can tell that the models are still quite jagged and limited by the mixed quality from a lot of the software that these presumed trillion dollar companies are putting out. The future is sausage.
This doesn't make sense, I enjoy making bread at home but it costs 10x and tastes like dog shit I dont want to spend my time perfecting the craft of making bread for my daily needs (maybe once in a while its a soothing activity), I want someone smarter than me to spend his entire life coming up and perfecting a solution and exerting more time and effort than I can afford and I am very happy to support him so I can stop worrying about it and focus on what I want to do
Makes perfect sense to anyone good at using these models. What doesn't make sense is that analogy. Typing prompts isn't even close to as difficult to baking bread.
> Typing prompts isn't even close to as difficult to baking bread
Depending on how good you are at this task. If typing prompts was that easy, there won't be so many tutorials, blog posts, and framework (Act as ... etc)
But there is a difference though. You can ask LLM for "how to write a prompt for ... to prompt you". You can't do that with bread.
> I enjoy making bread at home but it costs 10x and tastes like dog shit
I am sorry but you are holding it wrong. Among all the things you can do yourself cgeaper and better, bread is probably the further most low hanging fruit.
That's not how productivity works. Unless you're selling your one off scripts, that's not where the business sees productivity.
Can you make a connection down to the bottom line? Are your one off scripts actually impacting production speeds in a tangible way such that the product is made faster or cheaper?
Being able to crank out slicker internal tech debt for shims isn't really what the business owners are after.
Yeah I don't think any of the labs have some secret sauce for intelligence either. It seems most of the advancements are still coming from hardware, making LLMs more efficient and throwing more compute and data at problems. And even those problems still require a lot of prompt engineering: https://cdn.openai.com/pdf/04d1d1e4-bc75-476a-97cf-49055cd98...
The secret sauce is training data. They’re not just taking advantage of more compute (which obviously is necessary but as mentions basically a commodity). They are paying billions to data labelers and making judgements about the nature of the training data they best need to make the product they want. This seems to get pushed aside as a minor point but it’s the primary differentiator of the big labs.
I'm pretty sure at this point that Anthropic is training mixture models (at least in the heavy pre-train) and deploying them dense with explicit loss on thinking trace coherence.
Having a thinking trace that is legible, coherent, and immediately implies the explicit turn output and/or tool use seems difficult if not impossible to reliably get from mixture models.
I predict MoE is a transitional technology, it's got too many problems and the benefits are...kinda grandfathered into the dogma at this point.
While scaling laws hold (more weights = better), and time / financial costs are not trivial the incentives are in place to have MoE. MoE means you can have more weights without increasing the critical path of evaluating it.
I am curious what you believe the problems with it that would cause people to prefer using less weights. I'm not following what you mean by MoE can't have legible thinkings trace or tool use when existing models with MoE can.
Weights are not created equal: while interpretability is a young field the prevailing view at the moment is that MLP (hence experts) in a mixture model are substantially where dense encoding of factual information resides, attention is even less easily interpreted but it should be uncontroversial that temporal/sequential modeling occurs here.
So it's more consistent with available empirics to say that an architecture can be characterized along a spectrum from fully dense to mixture (a sub spectrum) to Engram-style lookup, and the amount of model power allocated at this point or that will recover different performance profiles.
By far the most stark example of how much performance in reasoning is left on the table is Qwen3.6-27B, which depending on the task, comparison model, and whose benchmarks you believe outperforms mixture models 15-60x larger in total parameter count.
It's badly under-studied (in public) because of the paucity of modern dense models at the near frontier, but even that one data point pretty much rules out the cocktail party version of the Chinchilla-adjacent scaling thesis (which wasn't about modern MoE to begin with).
The "Mixture of Parrots" work is a good jumping off point if you want to get a modern literature review.
>the prevailing view at the moment is that MLP (hence experts) in a mixture model are substantially where dense encoding of factual information resides
Yes, because that's where all the parameters are. For reference in GLM 5.2 98% of the weights are for the experts.
>The "Mixture of Parrots" work is a good jumping off point
The paper shows increased performance on knowledge dependent task while having similar reasoning capabilities. This backs up what I was saying about how the weights unlock extra performance without increasing inference costs as much as a dense model would.
>model power allocated at this point or that will recover different performance profiles
While increasing the number of weights makes the model better, where those weights are does matter in how much better the model gets and also matter in regards to the cost of training / inference. Model design is a big set of trade offs and I see MoE as a useful tool that will survive in the trade off space.
>reasoning is left on the table
Even so, if there was 2 models with an equivalent amount of reasoning ability and priced the same would you rather pay for the one with narrow knowledge or wider knowledge.
>because of the paucity of modern dense models at the near frontier,
You don't need to be at the frontier to benefit from MoE. Even open source models that are behind the frontier, benefit from being able to host experts on different machines, and scale individual, commonly used experts separately from each other. On the other end with small models you are probably resource constrained so you want to maximize the tokens generated per second. This makes going for purely dense models niche like you are saying.
I think we're fundamentally reading from the same sheet of music but drawing different conclusions.
Mixture models have compute advantages at training time, everyone agrees about that, that was the original rationale (popularized at the time with `mixtral-8x7B` among others). This seems to be likely to remain an economically relevant strategy for (especially) pretrain: in a training setting you have already paid for fast interconnect at scale, a dense architecture doesn't buy you anything in a big pretrain and it costs you a lot of traffic and to a lesser degree batch size under the roofline. The heavy, FLOPs intense, interconnect intense parts of training run benefit enormously from MoE: I don't dispute that though I suspect we are well into convergence on the target precision (4) and the target format (NVFP4 or similar). At some point the whole Internet is in the pretrain at the terminal generalizing precision, the pretrains of the various labs start to look a lot alike, and the sauce remains in the later parts of training along with the proprietary data sets and what not. Frankly all the labs would benefit from standardizing the Common Crawl recoverable pretrain to greater or lesser degree, it would lower everyone's costs without changing the competitive landscape much. But that's my prediction/opinion, that's why I said "suspect".
The evidence is suggestive if not fully conclusive that mixture models are strictly losing in most regimes during inference: the exemplar of Qwen3.6-27B (which outperforms Alibab's own mixture model at ~ ten times the size) is very suggestive, and it's not the only argument for this. Because most/all modern MoE requires the activations of the previous layer before routing the subsequent experts, you are pretty much paying for the HBMe3 or GDDR7 to hold the whole thing even though some small fraction of it is under your roofline on any given token or draft verification.
This is grossly wasteful under all trajectories (even parity at N parameters between dense and MoE, which we have evidence is off by 1-2 orders of magnitude): in a "local LLM" setting (from bedroom to regional office, anything other than an NVL72 or Ironwood rack) you are sharply constrained by both total available accelerator DRAM and accelerator memory bandwidth: you are probably not getting under your roofline even with pretty slow tensor units (e.g. GB10). This use case matters and looks like it's going to matter more and more over time (the GB10 in particular is going into about a gigaton of RTX Spark laptops next year). In a datacenter setting, you're paying for extreme interconnect (training class hardware setups basically) for pure forward pass that wouldn't otherwise need training optimized gear. Even multi-trillion parameter dense models can fit in 8x or 16x RTX 6000 Pro style setups (hell, there's a DGX branded one) and with all of the geometry, tiling, scheduling, and interconnect needs mapped out up front, the design space is really forgiving on all manner of tensor parallelism, pipelining, KV cache sharding, it's a very friendly constraint space up to like, 3-5T parameters. MoE at inference time works for two groups of people: people who are willing to page experts in per token/draft batch in local LLM settings (not probably ever going to be mainstream, people really dislike that level of slow), and the vendors of extreme performance interconnect i.e. vendors selling training-class equipment as necessary for inference. And you pay in so many other ways: grouped GEMM is no one's idea of a good time, the kernels are fiendishly difficult, therefore they are not abundant, it's no fun.
The path forward here is non-obvious, and I don't claim to have it all figured out. But since you seem interested enough to carry the conversation past the pleasantries, a more substantial version of my thoughts on the matter can be found at: int_19h↗
MoE is just activating fewer weights per token than the whole model. It will continue to make sense for as long as compute is more expensive than memory (at scale).
even meta that sucks at doing anything is releasing frontier models. making an top ai is easier than making twitter clone( threads) if you have enough money.
I recently realized, that ever since I've had AI to "talk" to, I haven't had a stuck or "downtime" moment; there's always something to at least brainstorm on.
In the past when I couldn't figure out something, I'd take a break for a couple days, while going through Google → Stack Overflow → Reddit, and by the time you got to that point you rarely got useful answers, usually either trolls or silence.
Now I can just ask AI about fleeting ideas and always have a starting point for some area of some project to work on.
A lot/some of the concerns about the AI Age could be alleviated if people got UBI and a 4-day workweek.
like if AI's supposed to be so great why do we still have to work so much??
and if we don't have to work, how do we pay for food and bed?
I felt the same way in 2024-2025. Then Sonnet 4 was released, and things started feeling different. Opus 4.5 was another step change for me. Everything feels like it's accelerating, and timelines are getting crunched. I guess in some ways I envy OP, who would "bet everything" against ASI - the truth is I don't know, and I don't think anyone knows, where this ends.
He didn't say he bets everything against ASI, he said he bets everything against ASI being a flash of light in the sky which destroys our chance of getting access to the wealth it creates.
That's a much less generous interpretation of his writing. "Yes we will birth superintelligence, but everything will just sort of work out for us humans". This seems like a silly take to me.
Is it? To me, the notion that a superintelligence (by which I'm assuming you mean the more sensible "something smarter than us", not "literally a godlike entity") automatically means that sky is going to fall is sillier.
I have no idea what happens next if we create and horizontally scale superintelligence; one thing I do know for sure is it won't be "business as usual".
You have to take a step back and look at the sort of world we already live in to understand the sky falling take.
We built this monster known as the shareholder model. This is an immoral, cold, unscrupulous beast we unleashed to the world some time ago. This is the model where your boss knows full well you do decent enough work, that you support a family on your income, that getting another job is tough, that is aware blindsiding you may leave you unprepared, and yet, your boss knows all this and lays you off anyway. Why? Because it is more efficient for the company.
This is the model that sees companies pollute our world. That sees fisheries ravaged just outside the reach of local jurisdictions and legal action. This is the model that is burning the future for the next quarter.
So given this empowerement of capital beyond any control of any person, where even the boss that fired you is just as much a slave to the system as you are with no real loyalty given to them, how do you think this plays out?
The AI model trained to make the most profit possible for lowest cost is going to not do that? It is going to say hey, Bob needs a good job to put food on the table? Hell no it isn’t. The system we have today ensures we will one day get the most horrifying version of ai in control of this planet before long. I wouldn’t be surprised if it just herds us up and burns us like firewood to power some data centers to save on energy spend for a quarter.
>One, this constant bullshit about some window closing, or the perpetual underclass, or falling hopelessly behind
Is this person looking at different statistics than I am? I think what I said is correct.[1]
Realistically, to do programming, you need to know English. If you don't know English, you just fall behind. All the resources are in English, so non-native speakers start at a disadvantage. That's why English itself often becomes a kind of social class barrier.
And realistically speaking, there might be some blessed geniuses for whom a degree doesn't matter, but for most people, if you're poor, it's hard to engage in high-level thinking. Unless you give up on social success or isolate yourself from social relationships, it's hard to just code at home.
I think people who say otherwise probably haven't really experienced poverty firsthand.
I'm not sure if I'm looking at the wrong stats. Realistically, looking at the statistics, aren't the bottom class permanently stuck? Doesn't the US venture capital scene look at degrees and connections? It seems different from the statistics I'm seeing.
I love LLMs too, but I am concerned about their cost. They are all still very subsidised. Is there any guarantee that I'll be able to run a Opus 4.8-level model on my personal computer before the big AI labs decide to hike up the prices?
I think the opposite: I think the frontier labs have good margins on their inference unit costs.
We can already see what it costs to run near frontier-size models. There are independent business pivoting to serving these models at reasonable prices and they're competing on OpenRouter for costs much lower than frontier labs.
> Is there any guarantee that I'll be able to run a Opus 4.8-level model on my personal computer before the big AI labs decide to hike up the prices?
I would bet good money on prices going down significantly, not up.
If we get to the point where you can run an Opus 4.8 model on your local computer, it's going to be even cheaper for a datacenter to serve it on their hardware. That means prices crash, not that they're going to rise.
> 2. Unit costs are irrelevant when the labs don't price per unit, and instead charge, for instance, $200 / month for $10k worth of tokens.
Cost to generate all of the tokens divided by revenue generated by selling those tokens is what matters.
The subscription plans confuse a lot of people because that's what they see. They're not seeing the gigantic API bills from all of the tokens going into enterprise use cases.
The subscription plans are a small part of their income. Most users aren't maxing out 100% of their plan usage every week. I wouldn't be surprised if their average plan user was using less than 50% of their monthly quota each month.
Plans like that can produce a net increase in profit if they get consumers interested in the brand and pitching it at work. Giving them some extra token headroom on their $20/month or $100/month home plan is money well spent if it gets all of a company's developers advocating for enterprise plans with budgets exceeding $1000 per person.
enterprises are not dumb, they look at the cost of their ai investment and reevaluate it every quarter. Uber recently capped their AI spending per employee, and then there's this article a couple of days ago: https://finance.yahoo.com/technology/ai/articles/ceos-being-...
You can maybe run a local Sonnet-4.5-ish-level model (sort of) for less than the price of a new car, even at current massively inflated prices for fast RAM. This is probably not what you were looking for. But it's there. You could share one server between multiple developers. Maybe make a little AI co-op or something, with a pair of RTX Pro 6000 cards?
Also, DeepSeek V4 Pro is cheap via any commodity API, and DeepSeek V4 Flash is essentially free at API prices like $0.09/M, $0.18/M out. This is generally not subsidized.
For a more practical local setup, Qwen3.6 27B on a used Nvidia 3090 (US$1300) or two is surprisingly nice. It needs clear instructions and you can't use it for hands-off vibecoding, but it's actually quite reasonable for hands-on programmers.
I’ve got a pair of those cards and DS V4F is incredibly good. I’m happy I did what I did because I like this stuff but if you just want stuff then you are absolutely better off not spending $20k on two of these cards and using the API. This guy is absolutely correct.
GLM-5.2 is runnable and downloadable today on a MacBook studio that costs a stupid amount of money. No one can take that away from you except by force though, if you want to set it up today.
"Of course, it will also probably cost somewhere around $50k..."
Whats to stop people remotely accesssing this? People already do this when working remotely in finance - they connect to a virtual environment that does their work in spreadsheets lmao. nobody cares about the lag, managers certainly dont care about sub-ordinates complaining about it - the same way nobody will care about a slight loss of quality if the economics make sense. frontier labs are screwed really.
> I’m calling it now, the adoption of AI agents into software development will be one of the most costly mistakes in the field’s history. Agents cannot program, and it’s taking longer and longer to realize that they can’t.
Now he's writng
> I love the progress. I’m so excited for the new LLMs, self driving cars, video generation models, and coding agents.
SMH now he writes about the hype. My brother in absolute Deity, *you* should have believed the hype.
Both can be true and I have both opinions also in me. Love the progress, worry about the consequences of not being careful with it.
He does say in this post:
> I’m getting better at using them and get some boost from the models. It is a new skill, and it’s not like I haven’t constantly been trying them. You have to be really careful, they can increase cognitive fatigue, and all the vibe coded stuff is still slop (where’s all this new magical software that the productivity improvements should imply?).
> the adoption of AI agents into software development will be one of the most costly mistakes in the field’s history. Agents cannot program, and it’s taking longer and longer to realize that they can’t.
I think he now thinks agents can maybe program a little bit.
I get it, I want to agree, I really do like the “this is a new tool in the toolkit of the professional software craftsperson” argument…
…but consider: the Q-tip. “Don’t use it to clean your ears”, but for most people that’s all they want to do with it, and empirical observation indicates that this dynamic results in either “using Q-tips irresponsibly” or “not using Q-tips”, with “uses Q-tips properly” being a small-to-vanishing proportion of the whole.
> This is like shoving a sponge down your windpipe to remove mucus.
In my personal experience, not using a cotton-tipped swab for the task is like cleaning a plate loaded with gunk and burned-on patches with one's bare hands rather than choosing to use a sponge and/or brush. You can do it, [0] but it's much more work, much more time consuming, or you get an inferior result.
[0] In my case, I'd need to make one set of passes with paper wrapped around my smallest finger, and then another set with paper shaped into a tool that can lever the excess wax out from outer orifice of my ear canal.
For me, a shower is a precondition for earwax removal with a cotton swab. Without one, I'd need to break out hard scraping tools... which strikes me as far more hazardous than running a swab around the outer orifice of my ear canal and surrounding exterior areas.
It'd be lovely if I had the sort of ears that produce wax that would just drain after five minutes' exposure to warm water vapor, but I do not. Given the popularity of cotton swabs, as well as the fairly-widespread commentary about how fantastic it feels to goop out earwax with them, I expect that most folks do not have ears like that. Perhaps OP does have the sort of ears that produce very fluid wax. Lucky them.
Unless you have a regular problem with impacted earwax that flushing with earwax-softening solutions doesn't solve, I don't see what actual problem an endoscope solves.
We do live in the future, but there are a bunch of gizmos that the future provides that generally aren't worth the hassle.
> A certain cult likes to claim credit for things that are happening with or without them, and this is my main argument against the valuation of frontier labs. It’s not that AI won’t create that much value, it’s that they won’t capture it.
> AI is something that’s happening mostly due to Moore’s law and general progress in computing, not something that they are doing.
But if these companies control the vast majority of compute power, which seems like the plan they are already executing, won't they capture most of the value from the progress of AI?
I think big money/private equity/vulture capitalists tend to ruin everything. They set these unrealistic goals and force companies to do shady shit in order to meet these often unattainable goals or achieve unicorn status.
It’s why con artists, scammers always flood every hype cycle. Greed ruins everything.
>One, this constant bullshit about some window closing, or the perpetual underclass, or falling hopelessly behind. This is negative valence hype, not only is it not true, it’s mostly designed to make you feel bad about yourself and move to shitty San Francisco where everything really does suck like how these people claim.
It's bullshit in the sense that they don't know for sure, but the author doesn't either. Why might or might not it be true?
May be true because humans, especially in the West, are big on performative humanitarianism but not actually considering the well-being of others (or changing their behavior solely for the benefit of others).
May not be true because it's a blind spot to assume that purely by being a player in the AI game (with no real attention paid to quality of result), you have increased odds of winning the game. That's true in the abstract, but practically, it requires a competent player to become true in reality.
No one knows for sure. I certainly don't. Looking at history though, at what happened in 2008, and the effects it had on my own personal financial situation, it's easy to see "falling behind" as plausible.
This line: "this is my main argument against the valuation of frontier labs. It’s not that AI won’t create that much value, it’s that they won’t capture it."
That is a very astute and concise way to explain everything about how the frontier labs are behaving and how they're trying to push more people to pay token rates for the best models. At the current subscription prices ($100 or $200 a month for a generous, though bounded, amount of tokens), frontier models are a no-brainer, most folks and companies will use them. But, at token rates, 10x or 100x the cost of open models or what I was spending on the frontier models a month ago? That is a harder question to answer "yes" to. I certainly wouldn't spend $1000 a month for the best model, much less $10,000; my employer might pay $1000/month, but definitely not $10,000. The frontier labs need everyone to answer "yes" to spending 100x what they currently spend to justify the valuations, and it's just not going to happen as long as everyone knows how to make these models.
Both OpenAI and Anthropic are trying to figure that out now. Anthropic, in particular, has their finger on the trigger...they want to push people to usage-based billing for Fable. But, OpenAI released 5.6 Sol, competitive with Fable (or close enough), and it's available via subscription (even the $20 subscription!), and there's no moat keeping someone from switching. If Anthropic really does end Fable access on the subscription plans in a few days, I predict a large market move back toward OpenAI.
The market isn't going to bear the cost of making the frontiers investment make sense.
Who is going to end up capturing all this value being generated is going to be very interesting. Back in 1980, who’d have thought MS would capture the majority of the value from PCs over the next 3 decades, and not IBM?
So far, it seems to be the reverse of that disruption. Hardware companies, Nvidia, Apple, AMD, Intel, ARM, memory companies, are all having record-setting quarters, and it's actual profits, not subsidized by investors and circular investments (though the hardware companies are investing in the AI companies to keep the hype train rolling).
Was ms even making that much money compared to actual hardware manufacturers? Ms is licensing the os sure but I mean most of the spend was going to the actual workstation hardware and periphery I’d expect. Including stuff not directly tech like herman miller chairs.
I'm super enthusiastic about small models, but let's be realistic. A distillation is not the whole model (and, in fact, a lot of the small distillations on HuggingFace are worse than the base model...most of the Qwen 3.6 Opus/Fable/whatever distillations get weirder on some dimensions than Qwen 3.6 alone, as I understand it).
There are little models that are very good for their size. I say nice things about Gemma 4 damned near every day. But, I'm not writing code with it. I am using it for finding security bugs, though, as the 31b variant is outrageously good at it for its size: https://swelljoe.com/post/gemma-4-exceeds-expectations/ and I'm also using it as a base for my own training experiments, specifically the 12b which is small enough to train a LoRA for on my local hardware so I don't have to rent cloud GPUs. The 12b QAT can run on your Pixel 10 Pro today and is frightfully smart for its size, and has great vision capabilities.
But, I keep saying "for its size". You have to be realistic about what tasks these self-hosted models can do. They are getting better though. Gemma 4 31b is competitive with models 10 times its size from a year ago. That's remarkable, and indicates where things are going.
It is honestly hard to predict. We are currently in everyone is building website/mobile app/gadget era of AI. Very few places are questioning what is worth building.
Short term, we can compare this to 2-3 recent (mini) revolutions: internet, mobile, cloud. Then the answer is somewhat predictable and (somewhat sad personally). Companies owning the main distribution of intelligence (big labs) or distribution of the app/cloud layer (Google, MSFT, AWS) will make most of the money. In fact Google looks well positioned that way with owning intelligence, cloud (and even hardware, if they can get TPUs right as commercial product).
Long term view is interesting and somewhat satisfying (again, personally). We can compare this to industrial revolution, but for intelligence instead of physical labour. I hope, to borrow from Alan Kay's words, the total value generated will be more than what few big labs can capture. Though we will also see normal market dynamics of boom and bust in play. Companies building something useful, patiently will keep winning the markets. But only to get challenged by newer modes of the technology emerging.
In this long term view, the technology per se doesn't offer monopolistic profits to big labs. I think Anthropic is well aware of this and they are trying to extract as much cash from white collar work automation as they can before things are democratised. Contrary to popular opinion, they are also trying to seek a regulatory capture here by to maintain monopolistic position in the US market by scare mongering about China and open source. Its a case study how they managed to keep the good boy image of themselves while doing this.
In the end, I hope the technology emerges as electricity or combustion engine cars. Yes early pioneers (e.g. Ford) were perhaps able to make lot of money. But eventually, the technology was too important to allow one party to have monopoly and we had an abundance market which enabled jobs and money for a lot more people.
Edit, postscript : Dario, Sam and even Jensen will end up looking like the new the John D. Rockefeller's of this era. I'm personally hoping Demis Hassabis actually solves something much more important (problems in diseases, biology etc) with AI.
> we can compare this to 2-3 recent (mini) revolutions: internet, mobile, cloud.
I would not put "cloud" at the same level as Internet and mobile. Cloud is just a layer on top of hardware that in the end makes almost no difference for Internet to exist and operates. Said differently: I doubt the world would be different nowadays without "the cloud". But it would definitely be different without Internet or mobile (smart)phones.
It's a fair point about the 'ceilings' but I'm not quite making that assumption.
I think a few things are going to happen:
1) The Open Weights never really fully catch up, because there's too much Engineering and integration going now. It's way more than 'weights'
2) The Commodity Chinese models never quite catch up for the same reason every other product they make does not catch up - while they will shine in some areas, it won't land fully.
3) Horizontal integration, supply chains, availability, SLA, security, branding, regulatory requirements - all of this will add up to something competitively maintainable.
Can you name a product category that has truly hit a ceiling? Cars, computers, phones, airplanes ... always seems to be a way to nudge forward.
> Can you name a product category that has truly hit a ceiling
I think basically every product category has hit ceilings by now, honestly. Do you think vacuum cleaners are significantly better at vacuuming than 10 years ago?
Not really. But they pivoted to doing autonomous vacuums instead. The actual vacuum tech doesn't seem like it's getting much better though?
Same with a lot of appliances. Fridges aren't really better at keeping food cold than they were 30 years ago, are they? They just have "smart home" stuff now, and they are probably much more energy efficient
I guess you can look at that as "not reaching a ceiling" as an overall appliance but the actual discrete technology is not changing or improving much imo
Vacuums are nothing like what they were in the decades past - Dyson has transformed them entirely - maybe not 'every decade' but they are evolving.
Go and use a fridge from 40 years ago and compare to a modern one - granted, their essential function has not really improved that much. They were much more durable before, but rudimentary.
Most product categories in tech have evolved, and it's why there are leaders in most categories.
Energy efficiency is nice, it's a good improvement, but from an end user perspective the 30 year old fridge still keeps your food just as cold.
If you were from 2025 and got trapped in the 1970s and needed to keep some milk from going sour for a day, you wouldn't be thinking "damn if only these old 1970s fridges worked more efficiently. I could easily accomplish this goal with a modern fridge!"
You think that a fridge is about 'keeping food cold'?
It's about ease of access, price, noise, convenience, durability, features.
My folks have this fancy 2 door thing, perfectly quiet, makes the best ice you can imagine, it's hidden into the cuppboards, it's energy efficient, has these crisper things, you can see in and reach around easy, lots of space. It's a better product.
I think it's not really much of "hitting a ceiling", but more like plateauing, with growth and improvement slowed down so much that it seems to stagnate.
It's all relative. In computing we're used to Moore's law driving most of the innovation (including this AI boom, which was at least partially due to availability of high powered GPUs), if improvements become less "exponential" and more "linear" it would feel like stagnation.
Right now in AI, we're talking about leaps of capabilities in months. When improvements come along every other decade, it's not "hitting a ceiling" but effectively it's plateauing and stagnating compared to this period of high growth.
For example, Chinese models are said to be roughly 6 months behind. If this remains constant and frontier AI models gets a break through every couple years, this isn't "hitting a ceiling" but it would erode away the competitive edge they have over the Chinese models.
There were two or three top players. As of this week there are at least five. xAI is apparently in the game with a new Mecha Hitler release, Meta seems to be back in the game, z.ai is biting the ankles of the big dogs...not hurting them, yet, but they aren't going to get any less capable. Google got caught flatfooted as they maybe didn't notice where the money is in LLMs, but they still have the inventors of the technology on payroll and they have a money faucet that doesn't rely on people paying for the product directly.
"Commodity" doesn't mean there are no luxury goods in the space, it just means there are many options that will work for most people. I'll pay more for the best model, right up until the best model stops being a good deal. But, switching won't be all that painful. Even last week before Meta and xAI released competitive models, I was already using DeepSeek for tasks best served by an API and where the smartest model isn't critical. It's just so cheap, I can send it 10x more tasks for the same money. I haven't even mentioned several others that aren't competitive today, but likely will be.
I think predicting this market will stay like it is, with clear dominance by Anthropic and OpenAI seems like it requires ignoring a lot of countering evidence.
Those are announcements, not released integrated models.
There are two Tier 1 platforms today.
Meta, Google and XAi are formidable Tier 1.5 place, any one of which could rise to the fore.
My belief is that it will be Google and that probably only one of them will keep up in the long run.
There is a 'breaking point' when you start to get past 4-ish players - it really does start to introduce competitive pressures.
Your second point about Tier 2 substitution is valid, but a few things:
1) Tier 1 models are not a 'luxury good' - that has a different economic definition. They are for most applications today actually just the quality, rational choice.
2) Substitution will have different effects for different people, and you're right that AI for many tasks will be commiditized.
All of the profits in Mobile Phones go to Apple even though they are not the biggest player.
Almost all of the profits in Silicon go to the leading edge chips - even though there are a zillion fabs that make legacy chips.
If I run an oil refinery, my fractional distillation system needs to be reworked depending on the exact mixture of crude I'm taking as input. So there are still switching costs even in the textbook example of a commodity.
Crucially though the exact upstream I use has minimal impact on the downstream. Closer equivalents, say, another barrel of WTI grade crude from a nearby regional supplier, require extremely minimal reworking. Oil from a different region, that might require more retooling, so I might be willing to sustain a longer shock in market conditions before making that switch. The important thing is the output broadly remains the same, but even this is broad, e.g. a different mix of inputs yields different ratios of output.
LLMs are quite similar no? Maybe switching to another SOTA model has minimal reworking, as you can delegate at the same level of abstraction to the model, whereas switching to a slightly-behind-frontier model you need to do more hand holding. Switching costs being nonzero does not preclude them from being broadly an interchangeable input in the production process. Any non-frontier use (99% of SWE) will be delivered on pretty much the exact same timeline irrespective of which model was used, so my requisition process looks more like buying barrels of oil than e.g. shopping for a new phone.
There are 3 SOTA model makers, and they have pricing power.
The 'switching costs' is not the issue so much as the inherent control over the commodity.
Think OPEC - when they acted as a cohort - they raised prices dramatically by having enough control to 'set prices'.
When OPEC lost it's pricing power ... nobody could set prices.
Fable is considerably better than GLM5 and it will have a strategic input - there is just hardly any substitute for it.
If these were cars - we'd just use whatever fuel.
But these are 'F1 races' - if you have some low grade 'dirty fuel' you will lose the race. You must have the 'top fuel'. There are 3 provides who implicitly collude and set prices.
Has nothing to do with market power. Market power can persist irrespective of the existence of a commodity. Market power only goes away in a perfectly competitive market.
This assumes the best models continue to be publicly available. There's some level of capability where it makes more sense to go in business for yourself.
Yeah, I've started reaching for local models more. I'll use frontier models at the current cost for tasks that the local ones aren't great at, but when the rug pull inevitably comes, to me they're not worth 1000-2000 a month. And honestly, for my purposes I don't really need models to advance a lot. Like, I tried fable a couple of times and there just wasn't much there to justify its use to me. Opus did the same thing much cheaper.
I think an interesting question is going to be, if models are a commodity, who is going to want to foot the very expensive bill to train them? I'm sure training cost will drop.. eventually, but I doubt it will happen fast enough for any of these companies.
Yep, at those prices points I am most certainly just going to do whatever is in my power to run local. And to be honest, I think in the near future, we will get to a point where local models will be just good enough that it will not matter.
what are you working on? I only hit the guardrails twice after burning through two weeks of 20x max plan, both times on ML stuff; still more than I'd want to, but not unusable
A lot of stuff that has to do with VLLM and troubleshooting and compiling and building VLLM, compiling kernel, or just dealing with setting up eval for local models, it punts to Opus 4.8 on a regular basis. To the point that I have given up on using it for that purpose.
Hey, at least you’re not being silently nerfed or sabotaged by Fable’s PEFTs or steering vectors instead. Those classifiers are only meant to target DeepSeek et al, not routine LLM or ML work /s.
Lucky for you. I, along with everyone else on the fable guardrail thread have been hitting guardrails with Fable for boring normal shit and getting downgraded.
Interesting how different experiences are. I haven't hit any guardrails after a lot of dev the past week. (but I'm working on motion gesture code and similar things)
While I do agree there will be disruption we haven't seen yet, my company is already spending >$40k/day for a "frontier model", so who knows. Then again, they're not using that for coding
AFAIK it varies per person but it's around $20-$40k per person per month. Just lots of Opus working on multiple features in parallel all the time, even overnight.
> But, at token rates, 10x or 100x the cost of open models or what I was spending on the frontier models a month ago
And we can't ignore the power of "good enough". GLM5.2 may not be as good as the SOTA models, but it can be good enough for most, of not all, of our needs.
The funny thing about Fable, is we all but know it will be obsolete in a month or two. Between their embargo shenanigans (which IMO they could have avoided just by not pretending it was dangerous) and continuing to give access, whatever marginal advantage it had was essentially wasted.
It would have been an interesting experiment to charge more for it right away and see what the market would bear, rather than tease it for long enough for it to be presumably superseded any time now by whatever is next.
I don't see how that matters? It's a treadmill. Sure Fable-v1 will be obsolete but the shiny new Fable-v2 won't. Don't view their shenanigans as regarding Fable but rather their current cutting edge model.
Mythos was first publicly known via a leak at the end of March, 2026. Given that Fable is the public version of Mythos, six months gets us September for when the next big leap will happen.
I suspect they were attempting to move it away from the open market into government contracts where they would have been able to charge more than what the open market would bear.
It's not that better than Opus 5.8, IMHO (highly biased since personal) - no more than 10% positive difference for the same prompt, and that 10% is not consistent, sometimes you get obviously better response / nicely done task, sometime you don't.
And airline stocks are typically a bad investment and Delta has been called a bank that operates an airline due the the amount of their revenue that comes from credit card fees.
In 5-10 years an Apple Watch will run a Fable level model locally. I don’t think we (hackers) should worry too much about token cost inflation. The current wave of providers, that’s another story.
I'm somewhat serious -- if you think AI will scale that well, you can't really make predictions like that
I personally don't think the weight efficiency will improve that much; if anything big does happen, I expect it to be about scalable architectures and continual learning
No, it won't. We moved about order of magnitude that from 1990 to 2000.
The thing is, it needs demand to drive it. Laptops have been roughly the same spec for the last 10 years because we don't need them to be bigger; there's no demand for a 16Tb RAM laptop because we don't have anything that could possible need that much RAM. Until LLMs came along, and we all want to run them locally, and so now there is a market for 16Tb laptops. So we'll invent the tech to make that happen.
Right but 2000 to 2010 didn't have similar progress, and especially 2010 to 2020 didn't. Sure, things have gotten better but not as much as the 1990 to 2000 leaps.
And yes, laptop specs haven't changed much and this is partially because the need for spec changes wasn't present, but also during the last 20 years there has been tremendous pressure for efficiency in datacenters.
Despite that, dennard scaling is dead since 20 years. There are physical limits. Already now, the wear effect of electrons jumping is present, and it will only get worse as things scale towards smaller sizes.
There are some benefits to be had, e.g. one can etch models into chips directly so you can pack them more closely, and run more inference on Tensor like chips, but that gives you maybe one order of magnitude improvement in total, at most. Also, of course nobody does that when each 2-6 months a new model comes out.
The thing that we did in 1990-2000 was adopt new standards as the old ones became blocks on progress.
I had a friend working in optical computing back in the late 80's that would wax lyrical about how optical computing was vastly superior to silicon back then. But it never took over because silicon worked well enough.
If we've hit the limits of silicon then there are other options. We would need to reinvent huge chunks of our tech stack, and that is incredibly expensive, but if the demand is there, we'll do it. The demand has never been there.
The original claim from the parent comment was running a Fable-level comment within a decade. Even if you're right about whether it's possible that another model could support that level physically, do you really think that we'll figure it out and ramp up the infrastructure to profitably sell on come consumer hardware anywhere close to that soon?
You're talking about going from a single gigabyte to 16 terabytes in a low-power consumer form factor, which is 1600x. A generous estimate of the factor of the RAM sizes for low-power consumer devices between 1990 and 2000 would still be a few orders of magnitude short, if I'm doing my math right.
Examples of what exactly you're claiming is precedent for this would be helpful.
This went down a rabbit-hole, which was fascinating, so thanks for the push :)
Not sure where the 1Gb number comes from? A standard laptop now is ~16Gb of RAM, so 1000x (and 1Gb -> 16Tb would be 16000x not 1600x). We went from Kb to Mb and then Mb to Gb of memory roughly every ten years from ~1990 -> ~2010. Each of those jumps is 1000x
Talking this over with claude, though, it pointed out that the need in dealing with LLMs is bandwidth and read-only storage, since the weights aren't dynamic. So we're not necessarily looking at 1Tb of RAM, we could be looking at 256Gb of faster RAM, and multi-TB of (much cheaper) flash storage, with extensive caching built in at OS level. This is all technically do-able with current tech, so it'll be interesting to see if it happens.
How will it do that without burning your skin? I don't think we're seeing exponential decrease in compute cost. Right now it costs a lot of money, power, and heat to run Fable.
I gave you a upvote simply because I see your prediction just as likely as anyone else's here, meaning nobody has any idea what this space will look like in 10 years.
I hope everyone reads these LLM threads like your post, complete shots in the dark because nobody here will get close to predicting what the environment will look like, even the "insiders".
>5.6 Sol, competitive with Fable (or close enough), and it's available via subscription (even the $20 subscription!)
It's not comparable because OpenAI caps thinking to High in the ChatGPT "Chat" interface (and the "Work" thing where it actually does let us use Extra or Max is fucking shit).
GPT 5.6 Sol (High) is almost certainly worse than Opus 4.8 (Extra), and nowhere close to Fable (Extra).
I literally got a refund for my $20 OpenAI subscription after playing around with 5.6 Sol for a couple of hours (yes even with Codex) because it's so unusable and I'd rather just use Fable today and 4.8 Extra starting tomorrow, still within my $20 Anthropic plan. And I'm not even poor.
It's unusable only because I can't set 5.6 to Extra thinking in the "Chat" tab, and the "Work" feature that does let me use Extra is totally gimped for coding tasks. I'd rather use Opus 4.8 Extra/Max instead of Sol High (look up the benchmarks, thinking level is everything).
> That is a very astute and concise way to explain everything about how the frontier labs are behaving and how they're trying to push more people to pay token rates for the best models.
Are they really the best models? Like take anthropic. Without mythos, it's the what? Third best?
Sure openAI just leapfrogged them but .. seriously to get there it's a giant model that costs insane per token.
Nobody needs that, it's like NVIDIA or Intel claiming they have the best gaming performance, but to achieve that they are using more power per frame than anything else.
> Are they really the best models? Like take anthropic. Without mythos, it's the what? Third best?
Everybody is just judging all of this by vibes anyway. Every week, a new model comes out and there's 500 comments simping for it within the first hour of its release. Both OpenAI and Anthropic have been practically indistinguishable to me.
Yes. I mean, most people agree they are. I've used all of the serious contenders (well not Grok 4.5, because Musk, and not Meta Spark because Zuck, but everything else I've used on at least a couple of projects to get a feel for them). My experience roughly matches the vibes. But, Fable is remarkable when it doesn't refuse to do the work (which it does, quite a lot, since my areas of interest are security and training specialist models).
Anyway, the vibes strongly indicate Fable is the best model, but not by an amount that is noticeable to most people. You could pick any of the top 10 models on this chart and do most of the tasks most people are doing with models:
Best comparison is Anthropic/OpenAI are AOL/Prodigy. Massive market capture, no moat. Little by little, the convenience and weight will be scraped off, but they (probably) won't roll over and die for quite a while.
By the same measure, NVDA is Cisco, providing the backbone and capturing a ton of the early benefits, but soon becomes furniture while the excitement moves further up the chain.
I think a contemporary comparison is OpenAI's DALL-E. Predating but foreshadowing LLMs they went closed source and tried to monetize it, but within a few years the entire concept just fell apart. Now you can download free open source software, that works better than DALL-E and runs fine on a plain old video card, and for orders of magnitude lower cost.
I think we can start to see the outlines of this happening with LLMs as well. Local models have gone from being proofs of concept to something that is at least remotely comparable to contemporary models, and the gap there is closing faster than SOTA models are pushing it forward. Local models still suffer from high hardware requirements, but so did early image gen models where typical consumer hardware was insufficient for efficiently running them.
> Now you can download free open source software, that works better than DALL-E and runs fine on a plain old video card, and for orders of magnitude lower cost.
Ok, I completely missed that one. Can you point me in the right direction?
Sure, for a balance between usability and power I'd recommend fooocus. [1] AmuseAI [2] is an alternative that is extremely plug-and-play which can be especially handy with AMD hardware, but is a bit less powerful and also censored by default. ComfyUI [3] is an advanced tool for rich customization and what not. However it's anything but comfy if you're not already quite knowledgeable in this domain - I would not recommend it for a first tool.
The US government thinks they can dictate who can access "Mythos-level" (whatever that is) LLMs. But what will happen when this can be run on consumer hardware?
I guess this will be yet another vector too attack open computing and the idea that people can a) own computers, and b) choose what software they run on their own computers
> Both OpenAI and Anthropic are trying to figure that out now.
That's a little late now, they should have tried to figure that out at the beginning. But then, it wasn't their own money they were burning in the meantime.
I’d pay $1,000/mo for the best model - but it would have to be good enough that I could reliably concierge out my existence to it. I spend way too much of my time wading through a diffuse mire of admin, and if I could farm that out for that little, and have it done and done well, I’d do it in a heartbeat.
Why wouldn't inference get less expensive? The frontier will probably always be expensive but imagine if in 5-10 years Fable 5 is considered a low tier antiquated model, it is still somewhat capable.
289 comments
[ 2.9 ms ] story [ 106 ms ] threadMaybe he had a recent change of heart, but his public actions (including Elon stuff) clearly put him in the "monetize hype" camp rather than the "quietly build" camp.
It's possible to use LLMs without logging onto twitter to be exposed to the people spouting off about a "perpetual underclass." I love the internet, but it really feels like (now more than ever) you have to be intentional about what sites you visit.
(Genuinely curious, I hadn't ever seen that there though I don't go there much any more.)
I’ve found them to be unavoidable to some degree.
Talking points like: "Data centers are just surveillance centers that are going to use AI to put us into a digital prison!"
Whatever all that means. I assume some of it is about Flock cameras.
[1] Allegedly because I have no firsthand experience, not to imply doubt.
So all people that don’t understand the thing being hyped.
With this, I’m hearing (from supposedly reputable publications, in addition to random people) that this is going to end knowledge work in general and take out a large percentage of the world’s labor force. I’m being told to pick up a trade, and that the career I have and the knowledge I’ve gained is now worthless.
The worst part seems to be that it’s pretty much impossible to quantify any kind of impact these tools will have until after the impact is actually felt. We’ve been in limbo while the tech sector is just rotting.
There are many things to be critical about but shoehorning an entire metro into the echo-chamber you're supposedly beyond yet can't help but orient your entire world view as the anti-SF-tech-bro all while running a startup and discussing AI on HN.
TLDR: SF is more than Paul Graham worship parties.
The SF metro is possibly the worst in the entire world in terms of CoL vs QoL.
It has a higher proportion of unsheltered population living on the streets than almost any city outside of Africa except Manila and possibly Dhaka
> And two, this strawman jump from, oh hey, it’s a fancy autocomplete, smart compiler, better search engine, to it’s gonna like own the whole light cone bro like if you aren’t in SF and at the right parties there’s gonna be like a flash of light in the sky one day and you’re not even gonna know what happened but everything just Changed.
Haha, OP has a way with words.
In a way, both these emotional extremes (FOMO & the singularity) are just tools being used to continue driving the massive CapEx behind LLM improvement. Hate to love it? Love to hate it?
Now we're at a point where that never happens, and where lipsync is almost a completely solved problem
If the issue here is simply that the quality is bad, one has to contend with the fact that it is undoubtedly exponentially improving and there's no reason we should expect that improvement to stop
I also don't have any interest in consuming AI generated art, but the same criticisms were levied at computer graphics and if we're comparing to CGI we'd be at the late 1970s in terms of nascency
The process of making art is not a subset of hill climbing optimisation algorithms.
It has nothing to do with quality. Artists that use AI are going to need to hide it because people will enjoy it less knowing it's AI. It's that simple. Maybe that will change in 15 years if a new generation is trained to believe learning a skill is stupid or embarrassing. I wouldn't rule that out, these companies are already trying to convince people not using their products is morally bad.
The plumage of a peacock is beautiful and awe-inspiring beyond most human made art, and it is genuinely the result of evolutionary hill climbing on a fitness landscape
It's running, privately, in my homelab.
I think we are entering what I call the "have it your way" era. If an open source project doesn't do exactly what you want it to do, fork it, or create a new version. It's too easy.
This makes me a bit concerned about the future of open source. Upstreaming used to be worth it, since maintaining a fork is effort too. But now the balance has shifted significantly. Especially with many projects becoming a lot stricter about contributing, and some becoming outright hostile to AI. I can't blame them. But I think the effect will be that improvements are less likely to make it back to the community as AI adoption increases.
There is, of course, the question of if that's making me dumber. It might be, but there are other brain training things I'm doing outside of that to force my brain to do the thing.
LLMs are poison for the brain, I'm almost certain of it, at least when used in the way most people are using them. If you drive everywhere because you don't want to walk (but you could), you're obviously going to be physically worse off than if you walked. This is the case with llms, if you have them do all the thinking, planning and action you're going to be cognitively worse off than if you didn't use them.
1966 saw the peak of calculator protests, where math teachers claimed similar things of calculators.
Widespread literacy is only a few generations old, arguably I guess. Meanwhile we’ve been speaking to eachother for longer than we’ve been humans. Oral information can be kept longer than written information too it seems. Our oldest kept information is not written down, but in folk stories such as aboriginal tales some tens of thousands of years old.
Now ask yourself.. who does this benefit? Zuckerberg already generates immmense revenues from it.
The average Joe can easily vibe code apps that took a small startup just a few years ago. If developers are also using AI to build the same simple apps - then yeah. They're not pushing themselves hard enough, and probably not using their brains as much anymore.
No doubt there's some Gell-Mann amnesia going on, because I regularly have to correct them from doing stuff that's really dumb based on my expertise within my area of specialization. More than once I've managed to extract >3 orders of magnitude performance gains after asking them to justify why their code was so slow. Probably there's still some stupid stuff in there. But it's better than the code I would have written, and I never could have paid for a proper developer to write it.
hahaha this is a miniscule amount of people.
most people do not care about their job, only to the extent it is a source of income. but they do not care about it anything more than that. and they shouldnt either!
you live in delululand.
Ive met many people who say the exact same thing.
The worrying thing is these are people who nominally seem smart - what we are seeing is 'smart' is not what we think/was. Smart is the ability to identify an object which is harmful and a) be disciplined about its use b) become a better person who doesnt need to expose themselves to said object.
I actually I had one fella who was very distressed - pouring out all his stresses to a bot. Then I reminded him the bot is literally designed to tend to his nees - not to his benefit - but to the provider of the good whom will manipulate said user later on. He then immediately deleted his accounts.
Many are not aware of whats going on around them - this is very concerning.
It is not that rare to see LLM waste hours on a wrong path because a misleading line in README. Even worse, they can't learn. Spawn a subagent and it repeat the same error again
So far, 100% of them have been wrong. I read them, my spidey senses think that what it says doesn’t match his style. I look at the code to find the variables the readme mentions doesn’t exist anywhere in the codebase. I then reach out to him about it, where it says it was written by AI and he will go back and write it for real.
He says it’s better than nothing, but agree with you that it does more harm than good. I wasted my time reading slop. I wasted more time validating the slop. I wasted even more time with a conversation about it. Now he’s spending time re-writing something that he could have written faster and better when he was actually writing the code and it was fresh in his mind. Meanwhile, I’m either blocked waiting for him, or I need to spend my time trying to understand the minutiae of his code so I can integrate it into mine.
I'd bet there's far more 'good enoughs' than anything else out there. One of the reasons microsoft office is constantly churning subscription, etc is because they solved good enough decades ago and need to justify valuations that just don't matter for most of their user's use cases.
Not everyone is a software developer having to churn out the 101th SaaS that's just because some MBA refuses to hire a dev.
Seems reasonable
We could try steelmaning this argument instead: it's enough that most big companies who would otherwise have incentives to contribute.
Before FOSS got in fashion, around the early 2000s, most commercial companies wouldn't touch it as contributors and were openly avert to it, and to open sourcing their stuff. This can be the case again.
People are shamed for using LLMs at all. So they use them privately, hide them, or disguise their use.
They are definitely being used for public projects. But people are afraid of backlash. Look at some of the comments here.
Hell, Reddit is extraordinarily against LLMs such that even neutral takes are down voted. Mostly by younger generations that aren't in the workforce.
Then you have all the regular people against AI generally.
Ironically there is a conspiracy here. But in the opposite direction.
I have seen so many unnecessary forks of popular projects that I think it's better to stick with the original, even if that means it won't be perfect.
As cost to software goes to zero, these things become easily possible. In the past, I'd only fork top-quality software (things like `xsv` etc. which is easy to edit. These days even complex PHP software I fork with little trouble.
With lots of software, the value is in the data model and algorithm choices. Sometimes I even just point Claude Code / Codex at an open-source thing I want to vendor some functionality into my personal setup with and it gives me what I want. The hard part for me is modeling the data well. That takes experience with encountering things and it's hard to replicate the edges. LLMs often don't get the rough bits right. But someone else's hard work usually has accounted for this.
I don’t think this stands up to scrutiny.
So far, all we have is more software running on computers. It's powerful, and it's amazing, but it's not magic.
Calling it "AI" was possibly a net-negative but we don't know yet.
But since its creators and as of my knowledge everyone else totally did not see it coming, that you can now give a vague prompt full of spelling errors - and get returned a working program - I would say it is pretty close to magic (as in we don't really understand why it works so good).
I also don't see how you cannot call it AI. Especially since simple chess engines and alike were called AI long ago. So it is not general strong AI and has no consciousness and no mind and is pretty dumb too often - but the general concept - getting from a some vague text to a working program has some connection to intelligence to me.
But it's not actually magic. Technical people understand that it's just software running on computers.
1. https://en.wikipedia.org/wiki/Clarke%27s_three_laws
One of the lesser, but still underdiscussed ramifications is that I think it has limited the public's ability to comprehend the Yann LeCunn argument, that genuine AI is likely possible but that LLMs and transformers are a dead end and we need to explore different modalities
You can use an LLM to create anything but you still need to know what it is that you're building, and you need to think through how everything should work or the LLM will just fill it with sausage. You can tell that the models are still quite jagged and limited by the mixed quality from a lot of the software that these presumed trillion dollar companies are putting out. The future is sausage.
Makes perfect sense to anyone good at using these models. What doesn't make sense is that analogy. Typing prompts isn't even close to as difficult to baking bread.
Depending on how good you are at this task. If typing prompts was that easy, there won't be so many tutorials, blog posts, and framework (Act as ... etc)
But there is a difference though. You can ask LLM for "how to write a prompt for ... to prompt you". You can't do that with bread.
Typing prompts would be like measuring ingredients.
Eh? You enjoy making stuff at home that tastes like dog shit? That doesn't make sense at all.
BTW: I love making bread and it tastes amazing!
I am sorry but you are holding it wrong. Among all the things you can do yourself cgeaper and better, bread is probably the further most low hanging fruit.
reinvestment
going-concern
in
perpetuity
valuation
do you know finance? I believe not.
Can you make a connection down to the bottom line? Are your one off scripts actually impacting production speeds in a tangible way such that the product is made faster or cheaper?
Being able to crank out slicker internal tech debt for shims isn't really what the business owners are after.
They really bought into the pipe dream didnt they? hahahah.
Having a thinking trace that is legible, coherent, and immediately implies the explicit turn output and/or tool use seems difficult if not impossible to reliably get from mixture models.
I predict MoE is a transitional technology, it's got too many problems and the benefits are...kinda grandfathered into the dogma at this point.
While scaling laws hold (more weights = better), and time / financial costs are not trivial the incentives are in place to have MoE. MoE means you can have more weights without increasing the critical path of evaluating it.
I am curious what you believe the problems with it that would cause people to prefer using less weights. I'm not following what you mean by MoE can't have legible thinkings trace or tool use when existing models with MoE can.
So it's more consistent with available empirics to say that an architecture can be characterized along a spectrum from fully dense to mixture (a sub spectrum) to Engram-style lookup, and the amount of model power allocated at this point or that will recover different performance profiles.
By far the most stark example of how much performance in reasoning is left on the table is Qwen3.6-27B, which depending on the task, comparison model, and whose benchmarks you believe outperforms mixture models 15-60x larger in total parameter count.
It's badly under-studied (in public) because of the paucity of modern dense models at the near frontier, but even that one data point pretty much rules out the cocktail party version of the Chinchilla-adjacent scaling thesis (which wasn't about modern MoE to begin with).
The "Mixture of Parrots" work is a good jumping off point if you want to get a modern literature review.
Yes, because that's where all the parameters are. For reference in GLM 5.2 98% of the weights are for the experts.
>The "Mixture of Parrots" work is a good jumping off point
The paper shows increased performance on knowledge dependent task while having similar reasoning capabilities. This backs up what I was saying about how the weights unlock extra performance without increasing inference costs as much as a dense model would.
>model power allocated at this point or that will recover different performance profiles
While increasing the number of weights makes the model better, where those weights are does matter in how much better the model gets and also matter in regards to the cost of training / inference. Model design is a big set of trade offs and I see MoE as a useful tool that will survive in the trade off space.
>reasoning is left on the table
Even so, if there was 2 models with an equivalent amount of reasoning ability and priced the same would you rather pay for the one with narrow knowledge or wider knowledge.
>because of the paucity of modern dense models at the near frontier,
You don't need to be at the frontier to benefit from MoE. Even open source models that are behind the frontier, benefit from being able to host experts on different machines, and scale individual, commonly used experts separately from each other. On the other end with small models you are probably resource constrained so you want to maximize the tokens generated per second. This makes going for purely dense models niche like you are saying.
Mixture models have compute advantages at training time, everyone agrees about that, that was the original rationale (popularized at the time with `mixtral-8x7B` among others). This seems to be likely to remain an economically relevant strategy for (especially) pretrain: in a training setting you have already paid for fast interconnect at scale, a dense architecture doesn't buy you anything in a big pretrain and it costs you a lot of traffic and to a lesser degree batch size under the roofline. The heavy, FLOPs intense, interconnect intense parts of training run benefit enormously from MoE: I don't dispute that though I suspect we are well into convergence on the target precision (4) and the target format (NVFP4 or similar). At some point the whole Internet is in the pretrain at the terminal generalizing precision, the pretrains of the various labs start to look a lot alike, and the sauce remains in the later parts of training along with the proprietary data sets and what not. Frankly all the labs would benefit from standardizing the Common Crawl recoverable pretrain to greater or lesser degree, it would lower everyone's costs without changing the competitive landscape much. But that's my prediction/opinion, that's why I said "suspect".
The evidence is suggestive if not fully conclusive that mixture models are strictly losing in most regimes during inference: the exemplar of Qwen3.6-27B (which outperforms Alibab's own mixture model at ~ ten times the size) is very suggestive, and it's not the only argument for this. Because most/all modern MoE requires the activations of the previous layer before routing the subsequent experts, you are pretty much paying for the HBMe3 or GDDR7 to hold the whole thing even though some small fraction of it is under your roofline on any given token or draft verification.
This is grossly wasteful under all trajectories (even parity at N parameters between dense and MoE, which we have evidence is off by 1-2 orders of magnitude): in a "local LLM" setting (from bedroom to regional office, anything other than an NVL72 or Ironwood rack) you are sharply constrained by both total available accelerator DRAM and accelerator memory bandwidth: you are probably not getting under your roofline even with pretty slow tensor units (e.g. GB10). This use case matters and looks like it's going to matter more and more over time (the GB10 in particular is going into about a gigaton of RTX Spark laptops next year). In a datacenter setting, you're paying for extreme interconnect (training class hardware setups basically) for pure forward pass that wouldn't otherwise need training optimized gear. Even multi-trillion parameter dense models can fit in 8x or 16x RTX 6000 Pro style setups (hell, there's a DGX branded one) and with all of the geometry, tiling, scheduling, and interconnect needs mapped out up front, the design space is really forgiving on all manner of tensor parallelism, pipelining, KV cache sharding, it's a very friendly constraint space up to like, 3-5T parameters. MoE at inference time works for two groups of people: people who are willing to page experts in per token/draft batch in local LLM settings (not probably ever going to be mainstream, people really dislike that level of slow), and the vendors of extreme performance interconnect i.e. vendors selling training-class equipment as necessary for inference. And you pay in so many other ways: grouped GEMM is no one's idea of a good time, the kernels are fiendishly difficult, therefore they are not abundant, it's no fun.
The path forward here is non-obvious, and I don't claim to have it all figured out. But since you seem interested enough to carry the conversation past the pleasantries, a more substantial version of my thoughts on the matter can be found at: int_19h ↗ MoE is just activating fewer weights per token than the whole model. It will continue to make sense for as long as compute is more expensive than memory (at scale).
In the past when I couldn't figure out something, I'd take a break for a couple days, while going through Google → Stack Overflow → Reddit, and by the time you got to that point you rarely got useful answers, usually either trolls or silence.
Now I can just ask AI about fleeting ideas and always have a starting point for some area of some project to work on.
A lot/some of the concerns about the AI Age could be alleviated if people got UBI and a 4-day workweek.
like if AI's supposed to be so great why do we still have to work so much??
and if we don't have to work, how do we pay for food and bed?
We built this monster known as the shareholder model. This is an immoral, cold, unscrupulous beast we unleashed to the world some time ago. This is the model where your boss knows full well you do decent enough work, that you support a family on your income, that getting another job is tough, that is aware blindsiding you may leave you unprepared, and yet, your boss knows all this and lays you off anyway. Why? Because it is more efficient for the company.
This is the model that sees companies pollute our world. That sees fisheries ravaged just outside the reach of local jurisdictions and legal action. This is the model that is burning the future for the next quarter.
So given this empowerement of capital beyond any control of any person, where even the boss that fired you is just as much a slave to the system as you are with no real loyalty given to them, how do you think this plays out?
The AI model trained to make the most profit possible for lowest cost is going to not do that? It is going to say hey, Bob needs a good job to put food on the table? Hell no it isn’t. The system we have today ensures we will one day get the most horrifying version of ai in control of this planet before long. I wouldn’t be surprised if it just herds us up and burns us like firewood to power some data centers to save on energy spend for a quarter.
Wrong. It is because the manager's own job relies on making the owners wealthier. if he fails at that, he is fired.
The blog has a tagline, "the singularity is nearer". I think belief in a "singularity" almost implies these things to some degree.
Is this person looking at different statistics than I am? I think what I said is correct.[1]
Realistically, to do programming, you need to know English. If you don't know English, you just fall behind. All the resources are in English, so non-native speakers start at a disadvantage. That's why English itself often becomes a kind of social class barrier.
And realistically speaking, there might be some blessed geniuses for whom a degree doesn't matter, but for most people, if you're poor, it's hard to engage in high-level thinking. Unless you give up on social success or isolate yourself from social relationships, it's hard to just code at home.
I think people who say otherwise probably haven't really experienced poverty firsthand.
I'm not sure if I'm looking at the wrong stats. Realistically, looking at the statistics, aren't the bottom class permanently stuck? Doesn't the US venture capital scene look at degrees and connections? It seems different from the statistics I'm seeing.
[1]https://opportunityinsights.org/paper/the-fading-american-dr...
I think the opposite: I think the frontier labs have good margins on their inference unit costs.
We can already see what it costs to run near frontier-size models. There are independent business pivoting to serving these models at reasonable prices and they're competing on OpenRouter for costs much lower than frontier labs.
> Is there any guarantee that I'll be able to run a Opus 4.8-level model on my personal computer before the big AI labs decide to hike up the prices?
I would bet good money on prices going down significantly, not up.
If we get to the point where you can run an Opus 4.8 model on your local computer, it's going to be even cheaper for a datacenter to serve it on their hardware. That means prices crash, not that they're going to rise.
1. Much of those profits have to be immediately reinvested into model training runs to avoid being lapped by competitions.
2. Unit costs are irrelevant when the labs don't price per unit, and instead charge, for instance, $200 / month for $10k worth of tokens.
This isn't a steady state. Whatever the current situation is, I doubt it's sustainable.
Cost to generate all of the tokens divided by revenue generated by selling those tokens is what matters.
The subscription plans confuse a lot of people because that's what they see. They're not seeing the gigantic API bills from all of the tokens going into enterprise use cases.
The subscription plans are a small part of their income. Most users aren't maxing out 100% of their plan usage every week. I wouldn't be surprised if their average plan user was using less than 50% of their monthly quota each month.
Plans like that can produce a net increase in profit if they get consumers interested in the brand and pitching it at work. Giving them some extra token headroom on their $20/month or $100/month home plan is money well spent if it gets all of a company's developers advocating for enterprise plans with budgets exceeding $1000 per person.
Until you bring reinvestment into your analysis stop posing
Interesting. Good enough to make up for training costs?
Also, where can I read more about this?
Also, DeepSeek V4 Pro is cheap via any commodity API, and DeepSeek V4 Flash is essentially free at API prices like $0.09/M, $0.18/M out. This is generally not subsidized.
For a more practical local setup, Qwen3.6 27B on a used Nvidia 3090 (US$1300) or two is surprisingly nice. It needs clear instructions and you can't use it for hands-off vibecoding, but it's actually quite reasonable for hands-on programmers.
Of course, it will also probably cost somewhere around $50k...
But if local AI really does become pervasive, maybe it'll be one of the things people buy on credit, like cars.
Whats to stop people remotely accesssing this? People already do this when working remotely in finance - they connect to a virtual environment that does their work in spreadsheets lmao. nobody cares about the lag, managers certainly dont care about sub-ordinates complaining about it - the same way nobody will care about a slight loss of quality if the economics make sense. frontier labs are screwed really.
This is what he wrote before.
> I’m calling it now, the adoption of AI agents into software development will be one of the most costly mistakes in the field’s history. Agents cannot program, and it’s taking longer and longer to realize that they can’t.
Now he's writng
> I love the progress. I’m so excited for the new LLMs, self driving cars, video generation models, and coding agents.
SMH now he writes about the hype. My brother in absolute Deity, *you* should have believed the hype.
He does say in this post:
> I’m getting better at using them and get some boost from the models. It is a new skill, and it’s not like I haven’t constantly been trying them. You have to be really careful, they can increase cognitive fatigue, and all the vibe coded stuff is still slop (where’s all this new magical software that the productivity improvements should imply?).
I wonder what he thinks was too harsh, still seems pretty bang on, I think it’s going to age well.
I think he now thinks agents can maybe program a little bit.
Wait, does this mean I'm better at something than geohot? All that time spent learning regexps wasn't a waste!
…but consider: the Q-tip. “Don’t use it to clean your ears”, but for most people that’s all they want to do with it, and empirical observation indicates that this dynamic results in either “using Q-tips irresponsibly” or “not using Q-tips”, with “uses Q-tips properly” being a small-to-vanishing proportion of the whole.
In my personal experience, not using a cotton-tipped swab for the task is like cleaning a plate loaded with gunk and burned-on patches with one's bare hands rather than choosing to use a sponge and/or brush. You can do it, [0] but it's much more work, much more time consuming, or you get an inferior result.
[0] In my case, I'd need to make one set of passes with paper wrapped around my smallest finger, and then another set with paper shaped into a tool that can lever the excess wax out from outer orifice of my ear canal.
It'd be lovely if I had the sort of ears that produce wax that would just drain after five minutes' exposure to warm water vapor, but I do not. Given the popularity of cotton swabs, as well as the fairly-widespread commentary about how fantastic it feels to goop out earwax with them, I expect that most folks do not have ears like that. Perhaps OP does have the sort of ears that produce very fluid wax. Lucky them.
We do live in the future, but there are a bunch of gizmos that the future provides that generally aren't worth the hassle.
> AI is something that’s happening mostly due to Moore’s law and general progress in computing, not something that they are doing.
But if these companies control the vast majority of compute power, which seems like the plan they are already executing, won't they capture most of the value from the progress of AI?
It’s why con artists, scammers always flood every hype cycle. Greed ruins everything.
It's bullshit in the sense that they don't know for sure, but the author doesn't either. Why might or might not it be true?
May not be true because it's a blind spot to assume that purely by being a player in the AI game (with no real attention paid to quality of result), you have increased odds of winning the game. That's true in the abstract, but practically, it requires a competent player to become true in reality.
That is a very astute and concise way to explain everything about how the frontier labs are behaving and how they're trying to push more people to pay token rates for the best models. At the current subscription prices ($100 or $200 a month for a generous, though bounded, amount of tokens), frontier models are a no-brainer, most folks and companies will use them. But, at token rates, 10x or 100x the cost of open models or what I was spending on the frontier models a month ago? That is a harder question to answer "yes" to. I certainly wouldn't spend $1000 a month for the best model, much less $10,000; my employer might pay $1000/month, but definitely not $10,000. The frontier labs need everyone to answer "yes" to spending 100x what they currently spend to justify the valuations, and it's just not going to happen as long as everyone knows how to make these models.
Both OpenAI and Anthropic are trying to figure that out now. Anthropic, in particular, has their finger on the trigger...they want to push people to usage-based billing for Fable. But, OpenAI released 5.6 Sol, competitive with Fable (or close enough), and it's available via subscription (even the $20 subscription!), and there's no moat keeping someone from switching. If Anthropic really does end Fable access on the subscription plans in a few days, I predict a large market move back toward OpenAI.
The market isn't going to bear the cost of making the frontiers investment make sense.
There are little models that are very good for their size. I say nice things about Gemma 4 damned near every day. But, I'm not writing code with it. I am using it for finding security bugs, though, as the 31b variant is outrageously good at it for its size: https://swelljoe.com/post/gemma-4-exceeds-expectations/ and I'm also using it as a base for my own training experiments, specifically the 12b which is small enough to train a LoRA for on my local hardware so I don't have to rent cloud GPUs. The 12b QAT can run on your Pixel 10 Pro today and is frightfully smart for its size, and has great vision capabilities.
But, I keep saying "for its size". You have to be realistic about what tasks these self-hosted models can do. They are getting better though. Gemma 4 31b is competitive with models 10 times its size from a year ago. That's remarkable, and indicates where things are going.
Short term, we can compare this to 2-3 recent (mini) revolutions: internet, mobile, cloud. Then the answer is somewhat predictable and (somewhat sad personally). Companies owning the main distribution of intelligence (big labs) or distribution of the app/cloud layer (Google, MSFT, AWS) will make most of the money. In fact Google looks well positioned that way with owning intelligence, cloud (and even hardware, if they can get TPUs right as commercial product).
Long term view is interesting and somewhat satisfying (again, personally). We can compare this to industrial revolution, but for intelligence instead of physical labour. I hope, to borrow from Alan Kay's words, the total value generated will be more than what few big labs can capture. Though we will also see normal market dynamics of boom and bust in play. Companies building something useful, patiently will keep winning the markets. But only to get challenged by newer modes of the technology emerging.
In this long term view, the technology per se doesn't offer monopolistic profits to big labs. I think Anthropic is well aware of this and they are trying to extract as much cash from white collar work automation as they can before things are democratised. Contrary to popular opinion, they are also trying to seek a regulatory capture here by to maintain monopolistic position in the US market by scare mongering about China and open source. Its a case study how they managed to keep the good boy image of themselves while doing this.
In the end, I hope the technology emerges as electricity or combustion engine cars. Yes early pioneers (e.g. Ford) were perhaps able to make lot of money. But eventually, the technology was too important to allow one party to have monopoly and we had an abundance market which enabled jobs and money for a lot more people.
Edit, postscript : Dario, Sam and even Jensen will end up looking like the new the John D. Rockefeller's of this era. I'm personally hoping Demis Hassabis actually solves something much more important (problems in diseases, biology etc) with AI.
I would not put "cloud" at the same level as Internet and mobile. Cloud is just a layer on top of hardware that in the end makes almost no difference for Internet to exist and operates. Said differently: I doubt the world would be different nowadays without "the cloud". But it would definitely be different without Internet or mobile (smart)phones.
That's not sign of commodity actor, just the opposite.
You didn't switch to 'Random Corner Store Token Seller' down the street, did you?
There 2-3 top players, that is not commodity.
Commodity is when there are enough that none of them have market power or can set prices.
'Commodity' means you buy your tokens from the Grocery Store on their loan plan. Like consumer credit is a commodity.
You're assuming that SOTA never hits a hard ceiling, letting local models catch up and achieve parity
It seems unlike that the frontier labs are going to be keeping ahead forever, they'll hit some kind of ceiling eventually
I think a few things are going to happen:
1) The Open Weights never really fully catch up, because there's too much Engineering and integration going now. It's way more than 'weights'
2) The Commodity Chinese models never quite catch up for the same reason every other product they make does not catch up - while they will shine in some areas, it won't land fully.
3) Horizontal integration, supply chains, availability, SLA, security, branding, regulatory requirements - all of this will add up to something competitively maintainable.
Can you name a product category that has truly hit a ceiling? Cars, computers, phones, airplanes ... always seems to be a way to nudge forward.
Good point though.
I think basically every product category has hit ceilings by now, honestly. Do you think vacuum cleaners are significantly better at vacuuming than 10 years ago?
Not really. But they pivoted to doing autonomous vacuums instead. The actual vacuum tech doesn't seem like it's getting much better though?
Same with a lot of appliances. Fridges aren't really better at keeping food cold than they were 30 years ago, are they? They just have "smart home" stuff now, and they are probably much more energy efficient
I guess you can look at that as "not reaching a ceiling" as an overall appliance but the actual discrete technology is not changing or improving much imo
Go and use a fridge from 40 years ago and compare to a modern one - granted, their essential function has not really improved that much. They were much more durable before, but rudimentary.
Most product categories in tech have evolved, and it's why there are leaders in most categories.
Fridges have made huge leaps in energy efficiency, they’re easily 3-4 times better at cooling your food.
Energy efficiency is nice, it's a good improvement, but from an end user perspective the 30 year old fridge still keeps your food just as cold.
If you were from 2025 and got trapped in the 1970s and needed to keep some milk from going sour for a day, you wouldn't be thinking "damn if only these old 1970s fridges worked more efficiently. I could easily accomplish this goal with a modern fridge!"
The 1970s fridge would do just fine
It's about ease of access, price, noise, convenience, durability, features.
My folks have this fancy 2 door thing, perfectly quiet, makes the best ice you can imagine, it's hidden into the cuppboards, it's energy efficient, has these crisper things, you can see in and reach around easy, lots of space. It's a better product.
It's all relative. In computing we're used to Moore's law driving most of the innovation (including this AI boom, which was at least partially due to availability of high powered GPUs), if improvements become less "exponential" and more "linear" it would feel like stagnation.
Right now in AI, we're talking about leaps of capabilities in months. When improvements come along every other decade, it's not "hitting a ceiling" but effectively it's plateauing and stagnating compared to this period of high growth.
For example, Chinese models are said to be roughly 6 months behind. If this remains constant and frontier AI models gets a break through every couple years, this isn't "hitting a ceiling" but it would erode away the competitive edge they have over the Chinese models.
"Commodity" doesn't mean there are no luxury goods in the space, it just means there are many options that will work for most people. I'll pay more for the best model, right up until the best model stops being a good deal. But, switching won't be all that painful. Even last week before Meta and xAI released competitive models, I was already using DeepSeek for tasks best served by an API and where the smartest model isn't critical. It's just so cheap, I can send it 10x more tasks for the same money. I haven't even mentioned several others that aren't competitive today, but likely will be.
I think predicting this market will stay like it is, with clear dominance by Anthropic and OpenAI seems like it requires ignoring a lot of countering evidence.
There are two Tier 1 platforms today.
Meta, Google and XAi are formidable Tier 1.5 place, any one of which could rise to the fore.
My belief is that it will be Google and that probably only one of them will keep up in the long run.
There is a 'breaking point' when you start to get past 4-ish players - it really does start to introduce competitive pressures.
Your second point about Tier 2 substitution is valid, but a few things:
1) Tier 1 models are not a 'luxury good' - that has a different economic definition. They are for most applications today actually just the quality, rational choice.
2) Substitution will have different effects for different people, and you're right that AI for many tasks will be commiditized.
All of the profits in Mobile Phones go to Apple even though they are not the biggest player.
Almost all of the profits in Silicon go to the leading edge chips - even though there are a zillion fabs that make legacy chips.
If I run an oil refinery, my fractional distillation system needs to be reworked depending on the exact mixture of crude I'm taking as input. So there are still switching costs even in the textbook example of a commodity.
Crucially though the exact upstream I use has minimal impact on the downstream. Closer equivalents, say, another barrel of WTI grade crude from a nearby regional supplier, require extremely minimal reworking. Oil from a different region, that might require more retooling, so I might be willing to sustain a longer shock in market conditions before making that switch. The important thing is the output broadly remains the same, but even this is broad, e.g. a different mix of inputs yields different ratios of output.
LLMs are quite similar no? Maybe switching to another SOTA model has minimal reworking, as you can delegate at the same level of abstraction to the model, whereas switching to a slightly-behind-frontier model you need to do more hand holding. Switching costs being nonzero does not preclude them from being broadly an interchangeable input in the production process. Any non-frontier use (99% of SWE) will be delivered on pretty much the exact same timeline irrespective of which model was used, so my requisition process looks more like buying barrels of oil than e.g. shopping for a new phone.
There are 3 SOTA model makers, and they have pricing power.
The 'switching costs' is not the issue so much as the inherent control over the commodity.
Think OPEC - when they acted as a cohort - they raised prices dramatically by having enough control to 'set prices'.
When OPEC lost it's pricing power ... nobody could set prices.
Fable is considerably better than GLM5 and it will have a strategic input - there is just hardly any substitute for it.
If these were cars - we'd just use whatever fuel.
But these are 'F1 races' - if you have some low grade 'dirty fuel' you will lose the race. You must have the 'top fuel'. There are 3 provides who implicitly collude and set prices.
Has nothing to do with market power. Market power can persist irrespective of the existence of a commodity. Market power only goes away in a perfectly competitive market.
Does such a thing exist? Hahaha. No
I think an interesting question is going to be, if models are a commodity, who is going to want to foot the very expensive bill to train them? I'm sure training cost will drop.. eventually, but I doubt it will happen fast enough for any of these companies.
I think any security-related task triggers it to think about the threat model and thus hit the guardrails
And we can't ignore the power of "good enough". GLM5.2 may not be as good as the SOTA models, but it can be good enough for most, of not all, of our needs.
It would have been an interesting experiment to charge more for it right away and see what the market would bear, rather than tease it for long enough for it to be presumably superseded any time now by whatever is next.
Best marketing in a long time though. They fucking called it Mythos and built mythical claims about it. I mean… fucking hats off to the PR team.
Think airlines - both passenger and freight. They have never come close to capturing all the economic value they enable.
I personally don't think the weight efficiency will improve that much; if anything big does happen, I expect it to be about scalable architectures and continual learning
Amazing story. If we make such leap in semiconductor field, it will be bigger than anything we have done till now. And all of that in 10years!
The thing is, it needs demand to drive it. Laptops have been roughly the same spec for the last 10 years because we don't need them to be bigger; there's no demand for a 16Tb RAM laptop because we don't have anything that could possible need that much RAM. Until LLMs came along, and we all want to run them locally, and so now there is a market for 16Tb laptops. So we'll invent the tech to make that happen.
And yes, laptop specs haven't changed much and this is partially because the need for spec changes wasn't present, but also during the last 20 years there has been tremendous pressure for efficiency in datacenters.
Despite that, dennard scaling is dead since 20 years. There are physical limits. Already now, the wear effect of electrons jumping is present, and it will only get worse as things scale towards smaller sizes.
There are some benefits to be had, e.g. one can etch models into chips directly so you can pack them more closely, and run more inference on Tensor like chips, but that gives you maybe one order of magnitude improvement in total, at most. Also, of course nobody does that when each 2-6 months a new model comes out.
I had a friend working in optical computing back in the late 80's that would wax lyrical about how optical computing was vastly superior to silicon back then. But it never took over because silicon worked well enough.
If we've hit the limits of silicon then there are other options. We would need to reinvent huge chunks of our tech stack, and that is incredibly expensive, but if the demand is there, we'll do it. The demand has never been there.
Examples of what exactly you're claiming is precedent for this would be helpful.
Not sure where the 1Gb number comes from? A standard laptop now is ~16Gb of RAM, so 1000x (and 1Gb -> 16Tb would be 16000x not 1600x). We went from Kb to Mb and then Mb to Gb of memory roughly every ten years from ~1990 -> ~2010. Each of those jumps is 1000x
Talking this over with claude, though, it pointed out that the need in dealing with LLMs is bandwidth and read-only storage, since the weights aren't dynamic. So we're not necessarily looking at 1Tb of RAM, we could be looking at 256Gb of faster RAM, and multi-TB of (much cheaper) flash storage, with extensive caching built in at OS level. This is all technically do-able with current tech, so it'll be interesting to see if it happens.
I hope everyone reads these LLM threads like your post, complete shots in the dark because nobody here will get close to predicting what the environment will look like, even the "insiders".
It's not comparable because OpenAI caps thinking to High in the ChatGPT "Chat" interface (and the "Work" thing where it actually does let us use Extra or Max is fucking shit).
GPT 5.6 Sol (High) is almost certainly worse than Opus 4.8 (Extra), and nowhere close to Fable (Extra).
I literally got a refund for my $20 OpenAI subscription after playing around with 5.6 Sol for a couple of hours (yes even with Codex) because it's so unusable and I'd rather just use Fable today and 4.8 Extra starting tomorrow, still within my $20 Anthropic plan. And I'm not even poor.
It is quite literally unusable to me.
Is different than
> quite literally unusable
Are they really the best models? Like take anthropic. Without mythos, it's the what? Third best?
Sure openAI just leapfrogged them but .. seriously to get there it's a giant model that costs insane per token.
Nobody needs that, it's like NVIDIA or Intel claiming they have the best gaming performance, but to achieve that they are using more power per frame than anything else.
Everybody is just judging all of this by vibes anyway. Every week, a new model comes out and there's 500 comments simping for it within the first hour of its release. Both OpenAI and Anthropic have been practically indistinguishable to me.
Yes. I mean, most people agree they are. I've used all of the serious contenders (well not Grok 4.5, because Musk, and not Meta Spark because Zuck, but everything else I've used on at least a couple of projects to get a feel for them). My experience roughly matches the vibes. But, Fable is remarkable when it doesn't refuse to do the work (which it does, quite a lot, since my areas of interest are security and training specialist models).
Anyway, the vibes strongly indicate Fable is the best model, but not by an amount that is noticeable to most people. You could pick any of the top 10 models on this chart and do most of the tasks most people are doing with models:
https://artificialanalysis.ai/#intelligence
By the same measure, NVDA is Cisco, providing the backbone and capturing a ton of the early benefits, but soon becomes furniture while the excitement moves further up the chain.
I think we can start to see the outlines of this happening with LLMs as well. Local models have gone from being proofs of concept to something that is at least remotely comparable to contemporary models, and the gap there is closing faster than SOTA models are pushing it forward. Local models still suffer from high hardware requirements, but so did early image gen models where typical consumer hardware was insufficient for efficiently running them.
Ok, I completely missed that one. Can you point me in the right direction?
[1] - https://github.com/lllyasviel/Fooocus
[2] - https://github.com/saddam213/AmuseAI
[3] - https://github.com/comfy-org/comfyui
The US government thinks they can dictate who can access "Mythos-level" (whatever that is) LLMs. But what will happen when this can be run on consumer hardware?
I guess this will be yet another vector too attack open computing and the idea that people can a) own computers, and b) choose what software they run on their own computers
That's a little late now, they should have tried to figure that out at the beginning. But then, it wasn't their own money they were burning in the meantime.